phnix.dev writing https://phnix.dev/blog/ Notes on reliable AI systems, local-first tools, evaluation, and research. en-ca Fine-tuning a protein model to find new antibiotics https://phnix.dev/posts/amp-discovery https://phnix.dev/posts/amp-discovery Mon, 01 Jun 2026 12:00:00 +0000 What a protein-model screening experiment taught me about leakage, class imbalance, cross-benchmark transfer, and honest evaluation. Honest results from a small GRPO lab https://phnix.dev/posts/rlvr-grpo-lab https://phnix.dev/posts/rlvr-grpo-lab Mon, 01 Jun 2026 12:00:00 +0000 What a small reasoning-model lab taught me about separating formatting failures from reasoning failures and keeping negative results. We spent months measuring whether our AI was getting better or worse https://phnix.dev/posts/rigr-agent-eval https://phnix.dev/posts/rigr-agent-eval Fri, 01 May 2026 12:00:00 +0000 Most teams ship AI agents without knowing if they actually work. Why I built Rigr to freeze baselines and catch regressions, and how it connects to measuring drift within a single run. Why I gave my coding assistant a local voice https://phnix.dev/posts/claude-voice https://phnix.dev/posts/claude-voice Fri, 01 May 2026 12:00:00 +0000 The product decisions behind claude-voice: local speech, automatic playback, interruption, synchronized highlighting, and a deliberately narrow scope. What Blackreach taught me about reliable browser agents https://phnix.dev/posts/how-blackreach-works https://phnix.dev/posts/how-blackreach-works Sun, 01 Mar 2026 12:00:00 +0000 Lessons from building a stateful browser agent: reduce noise, preserve progress, verify outcomes, and make uncertain results visible. What shared memory taught me about agent coordination https://phnix.dev/posts/velqua-mesh-architecture https://phnix.dev/posts/velqua-mesh-architecture Sun, 01 Mar 2026 12:00:00 +0000 Lessons from experimenting with shared memory across independent agents: provenance, relevance, ownership, correction, and deliberate handoffs. I built a GitHub dashboard for my terminal https://phnix.dev/posts/ghboard https://phnix.dev/posts/ghboard Sun, 01 Mar 2026 12:00:00 +0000 BubbleTea, Go, and why I never want to open a browser just to check my GitHub notifications again. Linear A, and why I want to take a crack at it https://phnix.dev/posts/linear-a https://phnix.dev/posts/linear-a Sun, 01 Mar 2026 12:00:00 +0000 70 years of failed phonetic matching on Linear A. Why the computational tools available now are different.